> ## Documentation Index
> Fetch the complete documentation index at: https://docs.servflow.io/llms.txt
> Use this file to discover all available pages before exploring further.
# Context fetching
> How an agent gathers what it needs before it answers, and when to use a step that always runs instead of a tool the agent chooses
Sometimes an agent needs to know something before it starts: which customer is writing, what is in stock, what a pull request changed. Context gathering is a set of steps that run first, every time, and hand what they find to the sub-agent's prompt. In a config file, a set of steps is called a context group.
## A step always runs. A tool runs if the agent decides to.
A context step and a [tool](/guides/add-a-tool) can do the same job, such as looking something up in a database. The difference is who decides.
| | A context step | A tool |
| ------------ | ----------------------------------- | ------------------------------------------- |
| When it runs | Every time, before the sub-agent | Only when the agent decides it needs it |
| Who decides | You | The agent |
| Good for | What every answer needs | What only some answers need |
| Example | Look up the customer who is writing | Search the orders when asked about an order |
Use a step for what you always need. It costs nothing in model decisions, and the agent cannot forget to do it. Use a tool for what you sometimes need.
## Using what a step found
Each step has a name, and the prompt reads the step's result by that name. In this **User prompt**, `customer` is a step that looked the customer up from the email address in the message:
```
The customer is {{ .customer.name }}, on the {{ .customer.plan }} plan.
Their message: {{ body "message" }}
```
You do not have to remember a step's name. In any prompt, **Insert a variable** lists every step under **Step & agent outputs** and writes the braces for you.
If a step found nothing, that part of the prompt is left empty, so write prompts that still read sensibly.
## One after another, or all at once
In order, which is how steps run unless you say otherwise, each step can use what the ones before it found. Look up the customer, then look up that customer's orders.
With **Run steps in parallel**, the steps start together. That is faster, and the steps cannot use each other's results. Use it for look-ups that do not depend on each other.
## When a step fails
If a step cannot finish, the run stops there and the sub-agent is not called. If an agent that gathers context goes quiet, check its steps first.
## Where it sits
Context gathering belongs to a sub-agent and runs just before it. On the canvas it is a node of its own, in front of the sub-agent. See [Agents](/concepts/agents) for the path a run takes.
## Set up context gathering in the dashboard
1. Open a sub-agent. Under **Flow**, find **Context gathering** and click **Configure context**. A context node appears in front of the sub-agent with one step in it, and its panel opens.
2. Under **Steps**, click **Configure step** on the step. Choose what the step does under **Type**, and fill in its fields.
3. To add another step, click **Add step**.
4. Optional: under **Execution**, turn on **Run steps in parallel**.
5. Click **Save**.
## Related
How entries, context gathering, and sub-agents fit together.
For what the agent should decide to do itself.
How a prompt reads the message and what steps found.
What a step can do.